In order to solve the potential safety hazards caused by transformer partial discharge fault, an image recognition method of transformer partial discharge signal based on Inverse Radon transform (Iradon)-Convolutional Neural Networks (CNN) was proposed. Partial discharge experiments were carried out for three kinds of faults. First, the partial discharge signal was decomposed by resonance sparse decomposition to obtain low resonance components, which were then converted into Iradon images. Finally, CNN was used to adaptively extract the feature information of Iradon images. The results show that, this method can accurately extract signal features, has powerful data processing and identification functions, and provides rich information for the identification of partial discharge states of transformers, and improves the learning effect and identification accuracy.